EG-ARSA: An Expert-Grounded Open Vision-Language Model for Road Safety Auditing (arXiv 2508.17630)

Forum topic · 小凯 · 2026-08-26

Summary

This paper introduces Expert-Grounded Distillation (EGD), an AI framework that transfers institutional road safety expertise into a compact vision-language model for scalable visual road safety auditing. Road traffic injuries are a major problem in low- and middle-income countries, where proactive auditing is limited by incomplete crash records, auditor shortages, and costly field inspections. In EGD, a teacher vision-language model is first calibrated against authoritative field audits; large-scale annotation proceeds only after the teacher reaches substantial agreement with expert risk assessments (Cohen's kappa = 0.74). The calibrated teacher then produces structured supervision, distilled via LoRA and a leak-free prompt into an 8B-parameter student model. The authors also release BD-ARSA, the first open expert-grounded visual road safety auditing dataset from Bangladesh with 21,947 image audit records, and the EG-ARSA model. Fine-tuning with expert grounding significantly outperforms zero-shot baselines, and the compact student surpasses its 31B teacher and Gemini-2.5-Flash in blind evaluations. Source: arXiv 2508.17630.

Paper Overview

Field: Computer Vision Authors: Md Thamed Bin Zaman Chowdhury, Moazzem Hossain Released: 2025-08-26 arXiv: 2508.17630

Abstract (Translated)

Road traffic injuries remain a major challenge in low- and middle-income countries, where proactive road safety auditing is limited by incomplete crash records, shortages of qualified auditors, and the high cost of large-scale field inspections. To address this problem, the authors propose Expert-Grounded Distillation (EGD), a novel AI framework that transfers institutional road safety expertise into a compact vision-language model for scalable visual road safety auditing.

The key innovation is a quantified expert-grounding stage: the teacher vision-language model is first calibrated against authoritative field audits, and large-scale annotation is permitted only after the teacher reaches substantial agreement with expert risk assessments (Cohen's kappa = 0.74). The calibrated teacher generates structured supervision signals, distilled via LoRA and a single leak-free prompt into an 8-billion-parameter student model.

The paper also releases:

BD-ARSA: the first open, expert-grounded visual road safety auditing dataset from Bangladesh, containing 21,947 image audit records
EG-ARSA: a model developed specifically for this task

Experiments show that expert-grounded fine-tuning significantly outperforms zero-shot baselines, and in blind evaluations the compact student model surpasses its 31-billion-parameter teacher as well as Gemini-2.5-Flash.

Key Points

Problem: Proactive road safety auditing in LMICs is constrained by incomplete crash records, auditor shortages, and expensive large-scale field inspections.
Method: A two-stage EGD pipeline — quantified teacher calibration against expert audits (kappa = 0.74 gate), followed by LoRA-based distillation into an 8B student with a leak-free prompt.
Dataset: BD-ARSA, 21,947 expert-grounded image audit records from Bangladesh, the first open dataset of its kind.
Results: Grounded fine-tuning beats zero-shot baselines; the 8B student outperforms the 31B teacher and Gemini-2.5-Flash in blind evaluation.

Paper link: https://arxiv.org/abs/2508.17630